ECCV 2026 Papers

Garment draping aims to fit clothing onto 3D human mod-els. Existing physics-based methods are computationally expensive andstruggle to integrate with general differentiable systems. In contrast, deeplearning-based methods often require extensive annotations and lack ex-plicit physical constraints, limiting their ability to model accurate details.We innovate by bridging physical models with deep learning and proposePhysDrape. It integrates three modules using forces as an intermediary:the force-driven GNN predicts forces at each node, the stretching solvermodels physical deformation, and the collision handler penalizes inter-penetration. Each module includes learnable parameters to fit underly-ing force propagation processes and physical properties, enabling gener-alization to unseen templates and control over properties like stiffness.PhysDrape enables end-to-end optimization, leveraging intrinsic garmentenergy for physically plausible deformation via self-supervised learning.Following the previous training and evaluation protocol on CLOTH3Ddataset, PhysDrape achieves lower energy scores, negligible interpene-tration, more realistic visualizations, and comparable time consumption.

Applying Multimodal Large Language Models (MLLMs) tolong-form video is challenged by the substantial computational cost ofprocessing long token sequences. Current training-free reduction meth-ods present a difficult trade-off between efficient and content diversity.Query-focused approaches are efficient but sacrifice content diversity,which is critical for general-purpose understanding and open-ended dia-logue. Conversely, methods that preserve diversity by operating on high-dimensional features often incur significant computational overhead.We introduce Perceptual Projection Pruning (P 3 ), a training-free frame-work that tries to address this trade-off. Our approach is based on the hy-pothesis that a low-dimensional yet informative representation can serveas an efficient proxy for diversity-aware visual token pruning. We firstproject high-dimensional features into a 3D perceptual space by comput-ing three complementary saliency axes: dynamic (motion), regional (con-textual contrast), and focal (local complexity). This representation thenserves as a foundation for a subsequent diversity-aware pruning strat-egy. Experiments across SOTA MLLMs show our method retains nearly95% of the vanilla model’s performance while pruning up to 75% of to-kens. Our work demonstrates that a carefully designed low-dimensionalrepresentation is highly effective for guiding diversity-aware visual tokenpruning in MLLMs.

Representation Autoencoders (RAEs) enable diffusion mod-els to operate in the feature spaces of pretrained visual encoders. How-ever, many off-the-shelf encoders are not optimized for faithful recon-struction and often discard fine-grained visual details. Finetuning theseencoders for image reconstruction can recover such details, but we showthat it also reduces the effective dimensionality of the resulting represen-tation space. We analyze how this altered geometry affects generation inhigh-dimensional feature spaces. Under this geometry, standard veloc-ity prediction in flow matching can require the model to fit orthogonalnoise directions outside the low-dimensional signal manifold, making op-timization inefficient. This motivates the clean data parameterization(x0 -prediction), which focuses learning on the underlying signal mani-fold. Across experiments with multiple reconstruction-finetuned featuresets, we show that x0 -prediction consistently improves text-to-image gen-eration performance.

Text-to-image diffusion models excel at generating images from nat-ural language prompts, yet fail to interpret numerical colors such as hex codes(#FF5733) and RGB values (rgb(255,87,51)). This limitation stems fromsubword tokenization, which fragments color codes into semantically meaning-less tokens that text encoders cannot map to coherent color representations. Wepresent NumColor, that enables precise numerical color control across multi-ple diffusion architectures. NumColor comprises two components: a Color To-ken Aggregator that detects color specifications regardless of tokenization, anda ColorBook containing 6,707 learnable embeddings that map colors to embed-ding space of text encoder in perceptually uniform CIE Lab space. We introducetwo auxiliary losses, directional alignment and interpolation consistency, to en-force geometric correspondence between Lab and embedding spaces, enablingsmooth color interpolation. To train the ColorBook, we construct NumColor-Data, a synthetic dataset of 500K rendered images with unambiguous color-to-pixel correspondence, eliminating the annotation ambiguity inherent in photo-graphic datasets. Although trained solely on FLUX, NumColor transfers zero-shot to SD3, SD3.5, PixArt-α, and PixArt-Σ without model-specific adaptation.NumColor improves numerical color accuracy by 4–9× across five models, whilesimultaneously improving color harmony scores by 10–30× on GenColorBenchbenchmark. The code is available at https://github.com/moatifbutt/numcolor/.

We introduce VidPair-Halluc, a new benchmark for evaluating video hallucination in large video models (LVMs) under rigorous and controlled conditions. Unlike previous benchmarks that primarily rely on text-based perturbations or adversarial questions while neglecting the consistency of visual backgrounds, VidPair-Halluc features video pairs with highly similar backgrounds but distinctly different foreground semantics, enabling precise attribution of model errors to genuine hallucination rather than background variation. The benchmark is constructed through PairFlow, a pipeline that leverages recent advances in text-to-image and video generation to systematically compose stories, generate coherent video clips, and assemble them into adversarial pairs. Covering both spatial and temporal reasoning across ten semantic aspects, VidPair-Halluc comprises 1K high-quality adversarial video pairs and 11K spatio-temporal QA pairs with control over background and foreground variations. Evaluations on mainstream LVMs show persistent difficulty with robust fine-grained video understanding in adversarial settings, and code and data are available at the project page.

Di!erentiable rendering has advanced novel view synthesis(NVS), yet applying it to real-world driving remains di"cult due tosparse capture viewpoints, dynamic objects, and limited multi-trajectorydata. We introduce the Multi-View Multi-Vehicle (MV2 ) datasetand benchmark for evaluating NVS models under large viewpoint changesin dynamic urban scenes. MV2 features synchronized captures from a car,scooter, and drone, each following distinct yet synchronized trajectories.Training NVS methods on one vehicle’s camera stream and testing on an-other enables evaluation under substantially larger viewpoint variationsthan existing single-trajectory datasets. All sequences are registered viaStructure-from-Motion and camera poses verified using manual pixel-level correspondence annotations, yielding 50 high-quality scenes with12000 images. Benchmarking recent NVS and camera pose estimationmethods shows that NVS performance degrades with increasing view-point disparity, and that feed-forward pose estimators notably lag behindoptimization-based approaches, highlighting MV2 as a rigorous testbedfor NVS in driving. The dataset, benchmark protocol, and project re-sources are available at https://mv2-dataset.github.io/.

High-resolution 3D radar data is scarce. Commodity mmWavesensors use small antenna arrays that limit angular resolution to several de-grees, and existing datasets provide only 2D range–azimuth maps or sparsepoint clouds rather than raw analog-to-digital converter (ADC) signals.Hardware scaling is expensive, synthetic-aperture scanning is impracticalat fleet scale, and learned synthesis methods are bottlenecked by the verydata shortage they aim to address. We present mmIR, an open-sourcedifferentiable frequency-modulated continuous-wave (FMCW) radar in-verse renderer that fits a physics-based forward model to real capturesand re-renders from dense virtual apertures to synthesize high-resolution3D radar data. Because radar resolution is too coarse to recover geom-etry directly, mmIR performs LiDAR-assisted inverse rendering: usingLiDAR-derived meshes as a geometric scaffold, mmIR optimizes per-vertex International Telecommunication Union (ITU) physics materials,vertex normals, and antenna beam patterns through end-to-end auto-matic differentiation of a phase-coherent multiple-input multiple-output(MIMO) forward model with multi-bounce propagation, polarization,and free-space diffraction. On seven outdoor and six indoor ColoRadarscenes, mmIR achieves 0.914 mean Pearson correlation on range–azimuthmaps versus 0.307 for Sionna-RT. Scenes trained on a cascaded imag-ing radar transfer to a co-located single-chip radar without re-training(0.554 correlation), and dense virtual arrays (100×100 elements) pro-duce single-frame 3D occupancy validated against LiDAR. Project page:https://mmwave-inverse-rendering.github.io/.

Online Video Instance Segmentation (VIS) relies on per-frame instance queries for causal detection, segmentation, and tempo-ral association. However, in realistic videos with occlusion, blur, andclutter, these predictions often become noisy—exhibiting false positives,missed instances, or inaccurate masks. Such noisy queries not only desta-bilize cross-frame association but also contaminate contrastive supervi-sion and propagate drift through memory-based matching. Despite ad-vances in recent online VIS methods, none explicitly identify or addressnoisy query contamination, leaving a fundamental weakness unexam-ined. We introduce MiNQVIS, the first noise-aware framework for on-line VIS that stabilizes both training and inference. Our approach con-sists of three complementary components: (1) query-level gating that fil-ters unreliable predictions using classification-guided thresholding andembedding-norm stabilization; (2) video-level prototypical contrastivelearning that aggregates only reliable matches into stable, noise-robustprototypes; (3) dual-memory association that fuses momentum embed-dings with history-best retrieval to prevent drift from corrupted frames.Experiments on YouTubeVIS 2019/2021 and OVIS show consistent im-provements over strong online VIS baselines, with large gains in challeng-ing scenes. Project Page: https://nothing898.github.io/MiNQVIS.

Depth estimation in multi-illumination scenes with multi-ple, spatially varying light sources remains a crucial yet less-exploredproblem. Illumination changes introduce shadows, specular highlights,and exposure shifts that distort local appearance cues, causing severedepth inconsistency or even failure. Existing depth foundation models,trained predominantly on uniformly lit data, degrade sharply under suchconditions. However, direct adaptation is challenging because groundtruth depth is typically limited for multi-illumination datasets, whilesynthetic relighting often incurs geometric distortions. To address thesechallenges, we propose LumiDepth, a framework that learns from multi-illumination RGB images. First, a Disagreement-Calibrated ProbabilisticPseudo Supervision (DCPS) module constructs high-quality pseudo la-bels while preserving diversity. Second, a Frequency-aware Consistencyand Distillation (FaCD) module improves cross-illumination stabilitywithout over-smoothing by enforcing low-frequency geometric consis-tency and distilling high-frequency structural details bi-directionally. Toenable systematic evaluation, we introduce ReMID, a real-world multi-illumination RGB-D benchmark, together with stability metrics thatquantify average and worst-case depth variation. Experiments acrossdiverse datasets demonstrate that LumiDepth achieves state-of-the-artoverall performance, markedly improving both consistency and accuracyby reducing depth variation by 30.2% and absolute relative error by24.8%. We further show our target-domain label-free design remains ef-fective for depth under other appearance shifts such as weather and sen-⋆sor noise.

Generating human motion from language in 3D scenes re-quires both semantic alignment with textual intent and physically plau-sible human-scene interactions. Existing methods struggle because im-plicit language-scene fusion entangles motion synthesis with scene rea-soning, while explicit grounding via fixed-vocabulary detectors or affor-dance maps lacks generalization and fails to capture full-body, temporallycoherent interactions. We address this representation gap by identifyingmutual distance as an effective interaction representation for modelingfull-body spatiotemporal human-scene relations, and propose MDNet,a two-stage diffusion framework that explicitly generates and groundssuch interactions. Since mutual distance encodes only relative geometryand is ambiguous for stochastic generation, we introduce VLM-based se-mantic anchors to provide absolute spatial grounding. We further modelmutual distance in the frequency domain to enhance temporal stabilitybefore synthesizing final motions. Experiments on HUMANISE demon-strate state-of-the-art semantic alignment, physical plausibility, and tem-poral coherence, with cross-dataset generalization to unseen scenes.

Real-time monocular 3D object detection remains challeng-ing due to severe depth ambiguity, viewpoint shifts, and the high compu-tational cost of 3D reasoning. Existing approaches either rely on LiDARor geometric priors to compensate for missing depth or sacrifice efficiencyto achieve competitive accuracy. We introduce LeAD-M3D, a monocu-lar 3D detector that achieves state-of-the-art accuracy and real-time in-ference without extra modalities. Our method is enabled by three keycomponents. Asymmetric Augmentation Denoising Distillation (A2D2)transfers geometric knowledge from a clean-image teacher to a MixUp-noised student via a quality- and importance-weighted depth-feature loss,enabling stronger depth reasoning without LiDAR. 3D-aware ConsistentMatching (CM3D ) improves prediction-to-ground truth assignment by in-tegrating 3D MGIoU into the matching score, yielding stable and precisesupervision. Finally, Confidence-Gated 3D Inference (CGI3D ) acceleratesinference by restricting expensive 3D regression to confident regions. To-gether, these contributions set a new Pareto frontier for monocular 3Ddetection: LeAD-M3D achieves state-of-the-art accuracy on KITTI andWaymo, and the best reported car AP on Rope3D, while running up to3.6 × faster than prior high-accuracy models (e.g., MonoDiff). LeAD-M3D demonstrates that high fidelity and real-time monocular 3D de-tection is simultaneously attainable, without LiDAR, stereo, or stronggeometric assumptions.

Controllable image generation methods, such as ControlNet,have demonstrated a remarkable capacity to introduce visual conditions(e.g., depth maps) to guide image generation. However, these methodsoften struggle with complex multi-instance scenes, frequently leading toattribute confusion among instances. While recent approaches attemptto mitigate this via manual instance labeling, such requirements arelabor-intensive. In this paper, we propose InstanceControl, a novel multi-instance controllable generation method that eliminates the need for in-stance labeling. We identify the primary bottleneck in existing meth-ods as the inability to accurately associate instance descriptions withtheir corresponding regions within visual conditions. To address this, weleverage the Vision-Language Model (VLM) to establish instance-levelcorrespondences between text prompts and visual conditions. Specifi-cally, the VLM automatically parses instance descriptions from the textprompts and simultaneously predicts instance masks based on the visualconditions. Furthermore, since the predicted masks may contain noise,we introduce an adaptive mask refinement strategy that dynamically re-fines these instance masks during the generation process. Extensive ex-periments demonstrate that our approach outperforms state-of-the-artmethods, achieving superior fidelity and precise instance-level control.The code and dataset will be made publicly available.

Functional 3D scene graphs offer a versatile and flexible rep-resentation for 3D scene understanding and robotic manipulation, de-fined by object nodes, interactive elements, and functional relationshipedges. However, their potential remains underexplored due to the limitedcoverage of existing benchmarks and the overly straightforward designof previous pipelines, which primarily focus on large-scale furniture butlack of hierarchical structures. Therefore, in this work, we extend thebenchmark coverage by introducing dense tabletop objects and explicitmulti-level functional relationships. This expansion introduces criticalchallenges involving small-scale, dense, and similar instances, with lack ofvisual anchoring in relational reasoning, instance confusion during cross-frame fusion, and attribution uncertainty under dynamic viewpoints. Toaddress these issues, we propose an open-vocabulary pipeline based on 2Dvisual grounding and 3D graph optimization. Specifically, we anchor fine-grained functional edges from 2D visual evidence, and associate nodesacross frames in 3D using multiple cues. Furthermore, edge associationis formulated as temporal graph optimization, integrating evidence ac-cumulation, entropy regularization, and temporal smoothing to robustlydetermine the functional connections of each node. Finally, global hierar-chy shaping is performed to recover the hierarchical graph structure. Ex-tensive experiments demonstrate that the proposed method can reliablyinfer functional 3D scene graphs in challenging real-world scenes, therebyfurther unlocking their potential for practical applications. Code is avail-able at https://github.com/Hbelief1998/HHOpenFunGraph-ECCV26.

Recent advancements in vision-language models (VLMs) haveleveraged large language models (LLMs) to achieve performance on parwith closed-source systems like GPT-4V. However, deploying these mod-els in real-world scenarios, particularly on resource-constrained devices,remains challenging due to their substantial computational demands.This has spurred interest in distilling knowledge from large VLMs intosmaller, more efficient counterparts. A key challenge arises here from thediversity of VLM architectures, which are built on different LLMs andemploy varying token types—differing in vocabulary size, token splits,Changing Teacher VLMs for Distillation MMMU70GenRecalMMB GPT-4oInternVL2.5-78B → InternVL2.5-8B60 Qwen2-VLAccuracy (%)MM-Vet InternVL2.550 MolmoMMMU40NVLM-72B → InternVL2.5-8B LLaVA-OneVisionInternVL2-76B → InternVL2.5-8BMMMU-Pro Qwen2-VL-72B → InternVL2.5-8BInternVL2.5-78B → InternVL2.5-8B300 20 40 60 80 1 2 4 7 8 … 72 78Accuracy (%) Model Size (B)Fig. 2: (Left) Comparison of the challenging benchmark performances, MMB [36], MM-Vet [63], MMMU [65], and MMMU-Pro [66] by changing large VLMs. The more pow-erful large VLMs we select, the greater the performance improvement we can achieve.(Right) Comparing the performance of the challenging benchmark: MMMU [65], withGenRecal and various VLMs across model sizes. Note that all the experiments in Fig. 1and Fig. 2 are conducted on the equal training dataset.and token index ordering. To address this challenge of limitation to aspecific VLM type, we present Generation after Recalibration (Gen-Recal), a general-purpose distillation framework for VLMs. GenRecalincorporates a Recalibrator that aligns and adapts feature representa-tions between heterogeneous VLMs, enabling effective knowledge trans-fer across different types of VLMs. Through extensive experiments onmultiple challenging benchmarks, we demonstrate that GenRecal signif-icantly improves baseline performances, eventually outperforming large-scale open- and closed-source VLMs.

This paper presents DINO-SLAM, a DINO-informed designstrategy to enhance implicit (Neural Radiance Field – NeRF) and explicitrepresentations (Gaussian Splatting – GS) in SLAM systems through themore comprehensive semantic understanding enabled by DINO. This lat-ter alone, however, lacks proper 3D geometry understanding, allowingonly for marginal improvements. Therefore, we rely on a Scene Geom-etry Encoder (SGE) to lift DINO features into geometry-aware DINOfeatures (geoDINO), to better understand those geometric relationshipsthat vanilla DINO features fail to capture. Building upon it, we pro-pose two foundational paradigms for NeRF and GS SLAM systems in-tegrating geoDINO features. Compared to state-of-the-art methods, ourDINO-informed pipelines achieve superior performance on the Replica,ScanNet, and TUM datasets.

Many challenges in scientific imaging involve solving ill-posedinverse problems, where the goal is to recover spatio-temporal fields fromindirect, noisy, and highly sparse measurements — often without accessto ground truth data or reliable simulators. To address this challengingscenario, we present NeuralDMD, an interpretable, untrained (per-instance) reconstruction framework that combines neural implicit rep-resentations with Dynamic Mode Decomposition (DMD) to reconstructcontinuous spatio-temporal dynamics directly from measurements. Neu-ralDMD parameterizes DMD modes as continuous neural fields, andimposes a low-rank linear dynamics prior with spectral time evolutionto enforce temporal continuity. This formulation enables both forecast-ing under sparsity, and yields interpretable modes and spectra. We findthat NeuralDMD outperforms baselines on a wide variety of tasks: fromweather data assimilation from sparse station observations to interfero-metric (Fourier domain) observations of Sagittarius A*, the black hole atthe center of our galaxy. Moreover, NeuralDMD remains stable whenextrapolating into the future. While this framework is most naturallysuited to linear dynamics, we show that it can be applied to nonlinearregimes, though with extrapolation performance that degrades with in-creasing nonlinearity. Together, these results show that NeuralDMDenables interpretable reconstruction and forecasting of spatio-temporaldynamics directly from sparse and indirect measurements without rely-ing on numerical simulators or training data.

Adapting video foundation models to volumetric medicalsegmentation by treating axial depth as a temporal dimension is attrac-tive, but the resulting slice propagation can be unstable when anatom-ical structures branch, disappear, or become poorly contrasted acrossslices. Static geometric prompts provide little guidance for newly ap-pearing or disconnected components, and SAM2-style streaming mem-ory may accumulate errors in these regions. We propose Di_x001B_uPrompt,a backbone-frozen, single-prompt framework that augments SAM2 withlatent trajectory priors for volumetric propagation. Given a user prompton a representative slice, an LLM-conditioned Neural ODE predicts acontinuous, patient-conditioned latent trajectory and converts it into apersistent prior memory. A dual-stream spatial gate then fuses this priormemory with the standard observation-driven working memory, allowingthe model to use image evidence when it is reliable and trajectory pri-ors when local propagation becomes ambiguous. Experiments on MSD,BTCV, and BraTS show improved promptable volumetric segmentationover SAM/SAM2-based baselines, especially on boundary-sensitive met-rics.

Image Manipulation Localization (IML) aims to achieve pixel-level localization of locally manipulated images. Its core challenge is howto distinguish the inconsistencies between authentic and tampered ar-eas. To this end, existing IML methods typically exploit patch-level andedge-level forensic cues for accurate localization. However, these methodsextract forged information at both levels with fixed parameters, leadingto interference from real information that is detrimental to localizationas more realistic manipulated textures are enhanced by the upgrading oftampering techniques. To tackle the above issue, this paper proposes aDisentangling and Gathering Forensic Cues method, called DG-Force,which explicitly disentangles and adaptively aggregates region and edgecues to preserve the most informative evidence. Specifically, we proposea Patch-based Forensic Disentangling (PFD) module and an Edge-basedForensic Disentangling (EFD) module to decompose forensic cues of dif-ferent granularities to explicitly suppress the imbalance between forensictraces in both patch-level and edge-level areas. In addition, a Multi-scale Forensic Transfer (MFT) module is further designed to aggregateand balance multi-granularity information through the interaction acrossdifferent scales for robust key inconsistencies in forensic traces discovery.Extensive experiments on multiple benchmarks demonstrate the superi-ority of our method in the image manipulation localization task.

Feed-forward 3D reconstruction models such as DUSt3R, VGGT, and Depth Anything 3 (DA3) are transformer–based foundation models that infer camera geometry and dense scene structure in a single forward pass. Trained at scale in a supervised fashion, they raise a central question: do these models build upon geometric principles akin to traditional multi–view pipelines, or do they primarily rely on learned priors arising from the large–scale training setup? We find that epipolar geometry emerges within the intermediate layers of all three models and is causally linked to correspondence patterns in attention heads. To study this, we perform a systematic analysis of their internal representations across three real–world datasets and a controlled synthetic dataset. We quantify geometric understanding by probing intermediate features, analyzing attention patterns to identify correspondence matching patterns, and performing targeted interventions at the attention level. Further, we assess the role of learned priors by applying challenging input–level perturbations, such as occlusions, scene ambiguities, and varying camera configurations, and compare them against classical multi–stage reconstruction pipelines.